Privacy-preserving multimodal fusion for Alzheimer's staging: A federated vision transformer framework with explainable AI.

Accurate, early-stage staging of Alzheimer's disease (AD) is critical for therapeutic intervention but is hampered by data privacy regulations, multimodal data heterogeneity, and the "black-box" nature of complex Artificial Intelligence (AI) models. To address these interconnected challenges, we introduce a novel, privacy-preserving federated learning framework for robust and interpretable AD staging. Our method integrates four clinically relevant modalities, 3D structural Magnetic Resonance Imaging (MRI), Image Biomarker Standardization Initiative (IBSI)-compliant radiomics, cognitive assessments, and U.S. Food and Drug Administration (FDA) cleared plasma biomarkers, within a parameter-efficient Swin-UNet transformer architecture. Key innovations include: (1) Low-Rank Adaptation (LoRA) and dynamic token gating, reducing communication overhead by 99% (to 140 KB/round) compared to standard federated averaging; (2) A hierarchical attention fusion mechanism that dynamically weights modalities based on predictive uncertainty; and (3) A fuzzy-rough hybrid explainability pipeline that synthesizes client-specific saliency maps into a consensus-driven, clinically coherent interpretation without sharing raw data. Evaluated on a stratified four-client federation derived from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort, our framework achieved a state-of-the-art balanced accuracy of 80.7%, closely approaching the centralized upper bound (82.1%) while demonstrating superior robustness to simulated domain shifts and consistent cross-site interpretability (Dice similarity: 0.84). This work establishes a foundational blueprint for the next generation of healthcare AI systems that are simultaneously accurate, efficient, privacy-preserving, and trustworthy, enabling scalable collaboration across distributed clinical networks.